GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling

A Model-Agnostic Framework for Detecting Anomalies and Errors in Multi-Agent Language Model Systems

Published

May 25, 2025

Authors: J. Zhou et al.
Published on Arxiv: 2025-05-25
Link: http://arxiv.org/abs/2505.19234v1
Institutions: King’s College London • Beijing Institute of Technology • Tsinghua University
Keywords: large language models, multi-agent systems, hallucination detection, error propagation, graph anomaly detection, temporal attributed graph, encoder-decoder architecture, information bottleneck, unsupervised learning, AI safety

Random Unsplash-style image

The rise of large language models (LLMs) has enabled the development of intelligent agents capable of complex, multi-turn dialogues, but collaboration between multiple LLM agents presents distinct safety challenges—including hallucination amplification and error propagation—that threaten reliability and security. Existing defense approaches struggle to address the dynamic nature of error propagation in such multi-agent systems, or require architectural changes, which limits their applicability.

To address these challenges, the article presents a comprehensive solution known as GUARDIAN, detailing its methodology and main contributions:

The article then reports experimental findings that demonstrate the effectiveness and robustness of the GUARDIAN framework:

In summary, the article concludes with several key takeaways that highlight the practical implications and future prospects of GUARDIAN: